Newer machine learning models

Now we will discuss several newer machine learning models that have emerged to resolve various limitations of the current models (for example, standard LSTMs). One such model is Phased LSTMs that allow us to pay attention to very specific events that happen in future during learning. Another model is Dilated RNNs (DRNNs), which provides a way to model complex dependencies present in the inputs. DRNNs also enable parallel computation of unrolled RNNs, compared with naïvely iterating through the unrolled RNNs.

Phased LSTM

Current LSTM networks have shown a remarkable performance in many of the sequential learning tasks. However, they are not well-suited for processing irregularly timed data, such as data provided by event-driven sensors. This is mainly because no matter whether an event is transpired or not, an LSTM's cell state and the hidden states are continuously updated. This behavior can cause the LSTM to ignore special events that might rarely or irregularly...